Snake Optimizer Improved Variational Mode Decomposition for Short-Term Prediction of Vehicle Charging Loads

Abstract

The rapid proliferation of electric vehicles (EVs) significantly impacts the power grid, necessitating effective forecasting of charging loads. For ultra short-term load prediction, this paper proposes a Snake Optimization (SO)-Variational Mode Decomposition (VMD)-Long Short-Term Memory (LSTM) algorithm trained by only the historical charging data. Before the prediction starts, the VMD method is utilized to minimize the data complexity, yielding several multiple Intrinsic Mode Functions (IMFs) that correspond to the charging load features at different time scales. The VMD parameters are automatically adjusted using the SO method, instead of the trial-and-error method, to trade off the prediction accuracy against computational overhead. Once the parameters of the VMD are determined, the same number of LSTM networks are employed to forecast the corresponding charging loads from these IMFs, with one LSTM for each IMF. Due to the VMD, IMFs with spanned center frequencies containing few irregularities make the prediction simple. These LSTM outcomes are then summed to obtain the overall load prediction. Experiments are carried out to show that the proposed parallel structure of multiple LSTM networks can achieve high prediction accuracy without requiring complex model structures. Our proposed algorithm outperforms the traditional prediction methods including Gate Recurrent Unit, Extreme Learning Machine, LSTM, and their combination with VMD, significantly reducing the Root Mean Square Error and the Mean Absolute Error by 30.1% and 32.9% in comparison with the optimal VMD-LSTM approach, and by 59.3% and 62.6% with respect to the baseline LSTM method.